Résumé:
The interest of collecting and exploring large amounts of data to extract valuable knowledge
has became critical to commercial companies and governmental organizations, which is the
motivation of data mining. The current tendency is that two or several organizations share
their datasets and give them as inputs to the process of data mining in order to have more
effective results. This raised a real problem of privacy that most of these data relate to
individuals and their personal information. The very active research area of privacy
preserving data mining aims to extract useful information from data coming from multiple
sources, while preserving these data against disclosure or loss. Clustering is also a more
exploratory data mining task which the aim is to classify items described by features into
groups, according to some similarities in a given context of application and poses the same
problem of privacy when data come from different sources. K-means is one of the algorithms
of clustering and the most widely used. Most of works in privacy preserving clustering are
developed on the k-means algorithm by applying the model of secure multi- party
computation. The ways in whish data are shared or distributed on these parties may be
different. The first solution of preserving privacy in k-means algorithm was proposed by
Jaideep Vaidya and Chris Clifton in 2003 on vertically partitioned dataset. Thus, approaches
allowing solving the problem on a vertical, horizontal and even arbitrary partitioned dataset
were proposed, but the preservation of privacy is still not complete. The major problem is to
reveal the minimum of information during the execution of the algorithm, especially in kmeans
iterations, which poses a real challenge for secure multi party computation. This paper
consists in drawing up a panorama of all works of preserving privacy in k-means clustering
algorithm, classifies the various approaches according to the used distribution dataset, while
presenting the weaknesses and strengths of each solution. The permanent growth of the data
and the need emerging to explore them requires a real thinking to effectively protect them,
especially that these data become increasingly individualized.
